Basal Ganglia neural network demonstrations, emergent v 7
These projects are downloadable for use with
the
emergent neural simulator.
Documentation is contained within
each project. It is strongly suggested that before diving into these
BG network
simulations, first familiarize
yourself with the emergent simulation package (both the software and
the
theoretical fundamentals, including neuronal and plasticity
equations). It will also be helpful to read the more detailed description of the computational models and associated
biology in the published
modeling
papers
(see Frank, 2005, 2006, Collins & Frank 2013, Wiecki & Frank 2013, and
Franklin & Frank 2015 for original model papers).
Note that these projects are not
available for the newest versions of emergent 8 (given the
various other changes that were
made). To run these, download the LTS emergent7.01 package on the
emergent site.
Start here - network
dynamics, gating, dopamine modulations of learning curves.
This project contains
a simplified Go/NoGo basal ganglia network and steps through the
roles of the different structures and their modulation by
dopamine, with basic replications of effects of Parkinson's and
medications on reinforcement learning in rich (mostly rewarding)
and lean (mostly punishing)
environments. New users should start here.
Probabilistic selection task .
This project is similar to above but implements the Probabilistic
Selection task with transfer
phase. Or use this project for
more detailed investigations.
Probabilistic selection (PS) task
simulations, tremor oscillations, various dopamine manipulations .
In depth simulations of recorded Go and NoGo striatal valuation signals and how these are
modulated by dopamine manipulations (depletion and medication
effects), including differential roles of D1 and D2
receptors, sensitivity to dopamine bursts and pauses, and
separable roles of
dopamine on both learning and choice incentive (expression of
learning). This simulation complements the above one by exposing
the neural mechanisms that generate the behavioral effects.c
Weather Prediction task
(probabilistic classification) simulations .
Simulates incremental learning of the challenging and now classical
Weather Prediction
task, and the effects of dopamine depletion on this learning. It
also includes the subthalamic nucleus (STN) and has a simple
demonstration of tremor-like oscillations that emerge with
dopamine depletion.
Task-set structured
learning, hierarchical corticostriatal circuit . From
Collins & Frank, 2013, Psychological Review. Includes
two-stage cascaded BG loop circuit enabling hierarchical control
of action selection and learning by generating task-set
structure, generalizable to novel situations. The model selects
among four different motor actions, and at the higher level,
three possible task-sets, and simultaneously learns to create
(or re-use) abstract task-sets while also learning the
particular response mappings given the selected task-set, using
pure reinforcement learning. This matlab
script can be used for more detailed analysis of model
output showing transfer, and here is
an example mat
file. Similarly, for more detailed analysis of a case in
which there is incentive to clustering task-sets around context
during initial learning, please use
this matlab
script. The computations of this model were linked to those
of a higher level "C-TS" (context task-set) model based on a
non-parametric Bayesian approach to clustering task-sets using a
Chinese Restaurant Process. Here is
a single zip file
including simulations from the C-TS model in matlab.
Computational model of inhibitory
control in prefrontal-basal ganglia circuits . From Wiecki &
Frank, 2013, Psychological Review. Includes
simulations of selective response inhibition tasks such as
antisaccade and Simon task, and the global response inhibition
stop-signal task. Captures various patterns of electrophysiology
observed in striatum, frontal eye fields, subthalamic nucleus,
superior colliculus, and elsewhere documented in such tasks, and
their relation to behavioral accuracy and RT distributions. The
script linked above includes a README file and Python code
which calls emergent neural software and analyzes the output.
Role of
cholinergic interneurons in adaptive reinforcement learning . From Franklin &
Frank, 2015, eLife. Shows how feedback circuit between medium
spiny neurons and tonically active neurons (TANs) can optimize
learning in an approximately Bayesian fashion and improve
performance in stochastic environments with reversal. TAN pauses
regulate population entropy across spiny neurons and are in turn
regulated by such entropy. Captures
effects of TAN lesions on reversal and is approximated by
Bayesian model of adaptive learning based on
uncertainty. Bayesian model scripts in Python
are here
and here
The models are implemented in
the emergent
neural simulator (Aisa et al., 2008) using a middle ground between biophysically detailed neurons
and highly abstract connectionist units. Physiological properties of
neuronal types in different BG nuclei are simulated by adjusting conductances and equilibrium
potentials of neurons. Synaptic weights are adjusted using pure
reinforcement learning as a function of changes in simulated dopamine
levels and their effects on striatal postsynaptic targets. (see
Frank, 2006 for a table of specific parameters and relation to BG
function).
|